Feature: compare_file_contents tool with semantic diffs

## Problem

When AI models review diffs, line-based unified diffs can be noisy and token-inefficient. Common scenarios where this hurts:

- **JSON/YAML reformatting**: A single value change plus auto-formatting creates a huge diff
- **Config file updates**: Version bumps or reordering keys produce misleading diffs
- **CSV/data files**: Row shifts make line-based diffs nearly unreadable

Models struggle to identify the *actual* change amid formatting noise, wasting context tokens and reducing comprehension accuracy.

## Proposed Solution

Add a new tool `compare_file_contents` that:

1. Takes two refs (base and head) plus a file path
2. For supported formats (JSON, YAML, CSV, TOML), produces a **semantic diff** showing only value changes
3. For unsupported formats, falls back to unified diff
4. Always shows the format used and whether fallback was applied

### Example: Semantic vs Line-based

**Line-based diff (noisy):**
```diff
-{"users":[{"id":1,"name":"Alice"},{"id":2,"name":"Bob"}]}
+{
+ "users": [
+ {"id": 1, "name": "Alice"},
+ {"id": 2, "name": "Bobby"}
+ ]
+}
```

**Semantic diff (clear):**
```
users[1].name: "Bob" → "Bobby"
```

### Tool Signature

```
compare_file_contents(
owner: string,
repo: string,
path: string,
base: string, // commit SHA, branch, or tag
head: string, // commit SHA, branch, or tag
)
```

### Use Cases

1. **Change verification**: Model edits a file, uses this tool to confirm only intended changes were made
2. **PR review**: Quickly understand what actually changed in config/data files
3. **Debugging**: Compare file across commits without formatting noise

## Implementation Notes

- Start behind a feature flag
- Semantic diff enabled by default for supported formats (no opt-out needed initially)
- Pure Go implementation using standard library JSON + yaml.v3
- Supported formats to start: JSON, YAML
- Future: CSV, TOML, other structured formats

## Why This Helps Models

- **Fewer tokens** = more room for reasoning
- **Unambiguous output** = clearer before/after semantics
- **Path notation** (e.g., `users[1].name`) is already familiar to models
- **Self-verification** = models can check their own edits efficiently